Let S = ( k = 1 ∑ 4 x k x k + 1 ) 2
Find the maximum of S when ∑ k = 1 5 x k 2 = 2 where x 1 , . . . , x 5 are positive real numbers.
You might enjoy this as well!
This section requires Javascript.
You are seeing this because something didn't load right. We suggest you, (a) try
refreshing the page, (b) enabling javascript if it is disabled on your browser and,
finally, (c)
loading the
non-javascript version of this page
. We're sorry about the hassle.
Very nicely explained! Thanks! (+1) This is certainly an impressive usage of those AM-GM inequalities.
As a challenge, can you apply this method to this problem ?
Great solution!!
Here is a solution for those who are familiar with quadratic forms .
The symmetric matric of the quadratic form q ( x 1 , . . . , x 5 ) = 2 ∑ k = 1 4 x k x k + 1 is A = ⎣ ⎢ ⎢ ⎢ ⎢ ⎡ 0 1 0 0 0 1 0 1 0 0 0 1 0 1 0 0 0 1 0 1 0 0 0 1 0 ⎦ ⎥ ⎥ ⎥ ⎥ ⎤ The characteristic polynomial is det ( A − x I 5 ) = − x ( x 2 − 1 ) ( x 2 − 3 ) , so that the largest eigenvalue of A is 3 . This is the maximal value of ∑ k = 1 4 x k x k + 1 when ∑ k = 1 5 x k 2 = 2 , with S = 3 .
The idea behind this technique is the following: Once we know the eigenvalues ± 3 , ± 1 , 0 of q , we can perform a change of variables and write the form as 3 v 1 2 + v 2 2 − v 3 2 − 3 v 4 2 . To maximize the value of q when ∑ k = 1 5 v k 2 = 2 we make v 1 2 = 2 and v k = 0 for k > 1 to obtain 2 3 . The sum we seek to maximize is 2 q .
(Equivalently, this solution can be phrased in the language of Lagrange multipliers.)
The idea behind this technique is the following: Once we know the eigenvalues ± 3 , ± 1 , 0 of q , we can perform a change of variables and write the form as 3 v 1 2 + v 2 2 − v 3 2 − 3 v 4 2 .
Wait.... so it's possible to avoid the ( A − 3 I 5 ) ⋅ ⟨ x 1 , x 2 , x 3 , x 4 , x 5 ⟩ T = 0 approach? I thought you previously said it couldn't be done?
Log in to reply
Indeed, you don't need the eigenvectors to find the extrema. You need the eigenvectors if you want to know where the extrema are attained.
What did I say could not be done?
Log in to reply
I asked : Thanks. It's ridiculously tedious (for me) to solve the Cayley Hamilton, det ( A − λ I 5 ) = 0 . How do you solve it manually (without the using of computer assistance)? Same goes with solving for B a = 0 . Is there a shortcut?
You replied : In principle, the eigenvalues are hard (perhaps even impossible) to find for matrices of size 5 x 5 and larger. In our example, the task is easy since 0 is a triple root so we are down to a quadratic polynomial. In principle, the eigenvectors are the easy part... just solving a linear system, by row reduction. The numbers can be messy, though.
Log in to reply
@Pi Han Goh – As Abel and Galois taught us, quintics cannot be solved in general , but many can, including the characteristic polynomial of our matrix here (for one thing, the matrix is singular).
Log in to reply
@Otto Bretscher – Ohhhh I misread what you wrote, haha! The way you wrote it made it deceptively simple.
Aren't quadratic forms and Lagrange multiplier two completely different stuff? How is it possible to "phrase" it in the language of LM?
Log in to reply
The linear system defining the Lagrange multipliers can be written as A v = 2 λ v , where A is the matrix of the quadratic form, so, it's exactly the same problem!
Wait... how do you know that the max of S is equivalent of searching for (max of S)^2? Because x 2 > x is not always true for 0 < x ≤ 2 .
Log in to reply
we have x > y > 0 iff x 2 > y 2 > 0 ... that's what we are using here.
Problem Loading...
Note Loading...
Set Loading...
First I'll show only what's necessary:
From AM-GM inequality, we have x 1 2 + 3 1 x 2 2 3 2 x 2 2 + 2 1 x 3 2 2 1 x 3 2 + 3 2 x 4 2 3 1 x 4 2 + x 5 2 ≥ 3 2 x 1 x 2 ≥ 3 2 x 2 x 3 ≥ 3 2 x 3 x 4 ≥ 3 2 x 4 x 5 Sum them up, we get x 1 2 + x 2 2 + x 3 2 + x 4 2 + x 5 2 ≥ 3 2 ( x 1 x 2 + x 2 x 3 + x 3 x 4 + x 4 x 5 ) It follows immediately that the maximum of S is 3 . It occurs when ( x 1 , x 2 , x 3 , x 4 , x 5 ) = ( 6 1 , 2 1 , 3 2 , 2 1 , 6 1 ) .
Here's how I know why I should divide them like that:
We first notice that the summation inside S is not cyclic. So we need to come up with a new kind of separation. Let 0 < k < 1 be an unknown real number. Since we have ∑ x i 2 = 2 , we should do something with it. We start out by x 1 2 + k x 2 2 ≥ 2 k x 1 x 2 . The number k x 2 2 is a fraction of x 2 2 that we want to combine with x 1 2 . It means that we use only k x 2 2 , not an entire x 2 2 . Now we have ( 1 − k ) x 2 2 left. Our next goal is to combine with a fraction of x 3 2 . The question arises that how much do I need to grab from an entire x 3 2 ? The answer is 1 − k k , why? Because we need a fraction of x 3 2 so that when we apply AM-GM, the right hand side is still 2 k x 2 x 3 , just like the first one. If we do like that, when we sum up, we would get 2 k ∑ x i x i + 1 . Now, we have ( 1 − k ) x 2 2 + ( 1 − k k ) x 3 2 ≥ 2 k x 2 x 3 . Similarly, we now have ( 1 − 1 − k k ) x 3 2 = ( 1 − k 1 − 2 k ) x 3 2 left. Iterate this method, we will get these inequalities:- ( 1 − k 1 − 2 k ) x 3 2 + ( 1 − 2 k k − k 2 ) x 4 2 ≥ 2 k x 3 x 4 , ( 1 − 2 k 1 − 3 k + k 2 ) x 4 2 + ( 1 − 3 k + k 2 k − 2 k 2 ) x 5 2 ≥ 2 k x 4 x 5 . Now, the right hand side is ready, but we still have a fraction of x 5 2 left, what should we do? The solution is take it all. Meaning that we choose k such that the fraction of x 5 2 in the last inequality is actually an entire x 5 2 . So we set 1 − 3 k + k 2 k − 2 k 2 = 1 . After solving this equation, we have k = 1 / 3 that is between zero and one. All we need to do after this is to check that it really works just like what I wrote above. That's the trick for question like this.